MOSS-IMAGE · COMFYUI · CHECKED OCTOBER 21, 2026
MOSS Image in ComfyUI
MOSS-Image runs in ComfyUI in two ways: with the native nodes that ComfyUI added on October 16, 2026, or with the optional MOSS-Image plugin from the official repo. This guide covers both, where every MOSS Image ComfyUI file goes, the default workflow settings, smaller builds for less VRAM, and the errors people hit most.
Based on the official ComfyUI guide and user guide. MossImage is an independent site, not part of OpenMOSS, and has not benchmarked these workflows itself.
Native nodes or the plugin?
Both load the same model. The native nodes are the recommended path; keep the plugin only if you need what it adds.
| Native ComfyUI nodes | MOSS-Image plugin | |
|---|---|---|
| Status | Built into ComfyUI (merged October 16, 2026) | Optional custom node from the official repo |
| ComfyUI version | Master or nightly from commit e087d8f6 | Any recent ComfyUI |
| Model files | Single files from Comfy-Org/MOSS-Image | A local Hugging Face / Diffusers folder |
| Loaders | Load Diffusion Model, CLIP Loader (moss_image), Load VAE | MOSS-Image HF Loader |
| Layout prompts | MOSS-Image Layout Prompt node | Not covered by the guide |
Set up MOSS Image with native nodes
- 1
Update ComfyUI
Pull ComfyUI master or nightly at commite087d8f6or later and reinstall itsrequirements.txt. As of October 19, 2026 the guide says release v0.41.3 does not include MOSS-Image yet. - 2
Download the ComfyUI files
Get the model, text encoder, VAE and example workflows from Comfy-Org/MOSS-Image on Hugging Face and put them in the folders below. - 3
Load a native workflow
Drag a native workflow into ComfyUI. It wires Load Diffusion Model, CLIP Loader with type moss_image, and Load VAE. Write your prompt and press Run. - 4
Change shift the native way
The built-in model already uses shift 5. To change it, add ModelSamplingAuraFlow with multiplier 1. Do not copy ModelSamplingSD3 from a plugin workflow: the native model uses a different timestep multiplier.
| File (Comfy-Org) | Put it in | Size |
|---|---|---|
| moss_image_sft_bf16 or moss_image_base_bf16 | models/diffusion_models | 10.2 GB |
| moss_image_sft_int8_convrot or base_int8_convrot | models/diffusion_models | 5.4 GB |
| qwen3.5_9b_bf16 | models/text_encoders | 19.3 GB |
| qwen3.5_9b_int8_convrot | models/text_encoders | 9.8 GB |
| qwen3.5_9b_w4a8_gptq | models/text_encoders | 7.3 GB |
| flux2-vae_small | models/vae | 0.25 GB |
SFT is tuned for finished images; Base is the raw pretrained model that most people use for LoRA training. The MOSS Image download guide compares them in more detail.
Set up the MOSS-Image plugin
Use the plugin when you want ComfyUI to read the official Base or SFT folder directly, or to keep running a workflow built on it.
- 1
Install the custom node
Copycomfyui/moss-image/from the official repo toComfyUI/custom_nodes/moss-image/and restart, or runpython comfyui/install.py --comfy-root /path/to/ComfyUIto link it. Your existing workflows are not overwritten. - 2
Open the example workflow
Dragmoss-image.jsoninto ComfyUI. For scripts, the repo also shipsmoss-image-api.json. - 3
Point the loader at your folder
In MOSS-Image HF Loader, setmodel_pathto the local folder that holdstransformer/,text_encoder/,tokenizer/andvae/. The loader reads local safetensors only; it cannot download from a Hub repo ID.
The repo also has a script pair that clones ComfyUI into a local folder with its own environment: bash comfyui/local/setup.sh then bash comfyui/local/run.sh.
Default MOSS Image workflow settings
| Setting | Default | Notes |
|---|---|---|
| Steps | 50 | The user guide suggests 50 or more for the best results. |
| CFG | 3 | Recommended range 3 to 6. Short prompts may need more than 3. |
| Shift | 5 | Native nodes: change it with ModelSamplingAuraFlow, multiplier 1. |
| Sampler | Euler, sgm_uniform | Euler is the recommended sampler. |
| Guidance skip | Layer 22 | Clear single_layers in Simple SLG to turn it off. |
| Canvas | 1024 × 1024 | Any ratio; sides must be multiples of 16. About 1 to 2 megapixels works best. |
| Text encoder | Up to 4,096 tokens | Long, detailed prompts are fine. |
The official negative prompt is a short tag list: nsfw, worst quality, bad anatomy, watermark, jpeg artifact. For prompt structure, English and Chinese text and line breaks, see the MOSS Image prompt guide. LoRAs with matching keys work with ComfyUI's built-in LoRA loader.
Run MOSS Image in ComfyUI with less VRAM
The full BF16 model and text encoder together are about 30 GB. Smaller builds trade some quality for size. They come from the community, not OpenMOSS:
| Build | Size | What you need |
|---|---|---|
| INT8 ConvRot (Comfy-Org) | Model 5.4 GB, text encoder 9.8 GB | Native nodes; files from Comfy-Org/MOSS-Image |
| Text encoder W4A8 | 7.3 GB | Native CLIP Loader, type moss_image |
| GGUF Q2_K to Q8_0 | 2.3 to 5.7 GB per model | ComfyUI-MOSS-GGUF plus city96 ComfyUI-GGUF |
A smaller file does not guarantee that MOSS-Image fits in 8 GB of VRAM: you still load the text encoder and VAE. GGUF files hold only the transformer and do not run in llama.cpp or LM Studio.
Common MOSS Image ComfyUI errors
| Problem | Fix |
|---|---|
| MOSS-Image nodes are missing | Your ComfyUI is older than commit e087d8f6. Update to master or nightly, then reinstall requirements.txt. |
| A plugin workflow fails after switching to native files | The loaders differ. Import a native workflow instead of swapping only the model loader. |
| Images look wrong with the native model | A ModelSamplingSD3 node copied from a plugin workflow uses the wrong multiplier. Use ModelSamplingAuraFlow with multiplier 1. |
| The HF Loader cannot find the model | Give model_path a local folder with transformer/, text_encoder/, tokenizer/ and vae/. It cannot download from a repo ID. |
| An error about the image size | Width and height must be positive multiples of 16. |
| Guidance patches stop working with TorchCompile | Run SLG or block patches with eager execution; TorchCompileModel filters out the guards they need. |
| Output differs from the Python pipeline | Expected: ComfyUI schedules guidance and timesteps differently, so the same seed is not a pixel match. |
MOSS Image ComfyUI questions
Does ComfyUI support MOSS Image natively?
Yes. According to the official ComfyUI guide, native support was merged into ComfyUI on October 16, 2026. You need ComfyUI master or nightly from commit e087d8f6 onward; the guide notes that the v0.41.3 release did not include it as of October 19, 2026.
Which MOSS Image files do I need for ComfyUI?
Three parts: the diffusion model (Base or SFT, in BF16 or INT8), the Qwen 3.5 9B text encoder and the small Flux 2 VAE. The Comfy-Org repackage on Hugging Face has them as single files for models/diffusion_models, models/text_encoders and models/vae.
Should I use the native nodes or the MOSS-Image plugin?
Use the native nodes. The plugin is optional and only useful if you want to load a Hugging Face or Diffusers folder directly, or keep using an older plugin workflow. The two use different loaders, so import a native workflow instead of swapping one loader.
What are the default MOSS Image settings in ComfyUI?
The official workflow uses 50 steps, CFG 3, shift 5, the Euler sampler with the sgm_uniform scheduler, negative-branch skipping at layer 22 and a 1024×1024 canvas. Width and height must be multiples of 16.
Will ComfyUI give the same image as the official Python pipeline?
Not exactly. The guide explains that ComfyUI measures guidance intervals in sigma percentages and its scheduler does not snap to the same grid, so the same seed and settings can still produce a different image.
Can I run MOSS Image in ComfyUI on a smaller GPU?
Community INT8 and GGUF builds make the files smaller, but a smaller file does not guarantee it fits in 8 GB of VRAM, and you still need the text encoder and VAE. GGUF files need the ComfyUI-MOSS-GGUF bridge plus city96 ComfyUI-GGUF.
No GPU? Run MOSS-Image online
Skip the setup and generate with MOSS-Image-SFT on MossImage, from 4 credits at 0.5K and 10 at 1K. New accounts get 20 free credits.